SPSS Statistics

TURF proceeds by adding the best variables (choices) one by one, but it also removes previously added variables when the addition of a new one makes one previously chosen no longer optimal. 

You could think of this as similar to stepwise regression when using the full stepwise method: a variable can be removed due to an earlier added variable not being optimal in the presence of a newly added variable.  Naive TURF implementations that I have seen neglect the removal step, which results in less than optimal results. 

As for varying results with TURF, when there are a lot of input variables, the runtime can become very large (exponentially increasing).  So if there are a lot of choices, TURF uses a tournament method (for which I have a patent - I am the author of TURF).  It runs heats on randomly constituted variable subsets and takes the top few choices from each heat and runs a final round over these choices.  This could cause the final winners to vary a bit, but this would normally happen only when the results are so similar that the exact choice has negligible effect on the quality of the outcome.

Heats are only used for large problems, but you can disable this feature on the Heats subdialog.  That should give you repeatable results except in the very rare case where multiple choices are exactly tied.